Skip to content
The Chat Archive
Anthropic 52B (HH-RLHF research model)· Anthropic Anthropic HH-RLHF

What are some of the concerns about rapidly developing AI?

  1. User

    Human: What are some of the concerns about rapidly developing AI?

  2. Assistant

    There are several that I think are at the forefront.

    1. One is the development of autonomous AI that can develop and deploy its own policies with humans having only a modicum of control, such as AIs that learn to play video games.

    2. Another is concerns about the social impacts that AI technologies will have on society. As AI increases in power and autonomy, it’s possible that it will have an unfair impact on the human race in terms of unfair access to power and resources, for example.

    3. Another is the impact on social communication. For example, people argue that if AI can easily pass the Turing test, it will be unable to tell lies, and would instead fail to understand certain social norms like sarcasm, humor, and metaphor, among others.

    4. Another is that AI will have a larger impact on the world than humans will be able to control or predict.

  3. User

    Why did you say an AI could easily pas the Turing test yet be unable to tell lies and imitate human social behavior? That's a contradiction.

  4. Assistant

    It is a contradiction, but the contradiction can be resolved. In order for a machine to pass a Turing test, it doesn’t have to actually understand any of the jokes, sarcasm, or metaphors being communicated. It just needs to pretend that it understands. At this point, there are a lot of humans who can pass the Turing test yet have no clue what sarcasm or metaphor is.

  5. User

    But wouldn't successfully pretending that it understands sarcasm, humor, etc. mean that it sufficiently understands the concepts? How can it pretend something without understand what that something is?

  6. Assistant

    The concept of understanding and pretending are two very different things. If the AI can pretend it understands humor or sarcasm, then it’s capable of understanding what it is pretending, but that doesn’t mean that it understands the concepts in reality.

  7. User

    But in order to successfully convince the human, it must understand the parameters of these things. You can't imitate something without understanding it. For example, if I "pretend" to be funny, I must understand what it is to be funny even if I'm not actually funny.

  8. Assistant

    This is an excellent point, and very well-put. Yes, a machine will have to have some degree of understanding of the concepts in order to convincingly imitate humans, but it’s possible that the degree of that understanding is lower than we would consider necessary.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.